80 Chapter 4. Fault Tree inference using Multi-Objective Evolutionary Algorithms and Confusion Matrix-based metrics gates. Table 4.2 outlines for each case study the number of Basic Events (|BEs|), FT Size (|F|), failure dataset size (|D|), total number of MCSs (|CD|), and the number of all FTs across generations. The latter is further discussed in Section 4.5.1. Implementation. The implementation of FT-MOEA-CM, available online∗, is complemented by a dedicated database server designed for storing and processing data produced by FT-MOEA-CM. This server, developed in GO, employs a MySQL database and can be accessed online†. Generation of failure dataset. Access to real-life failure data is typically very limited. Instead, we evaluate our approach on synthetic failure datasets which are generated from realistic reliability models. We consider existing FTs from the literature as ground truth, see Table 4.2. For each FT, we generate a synthetic failure dataset Dby evaluating all the unique combinations of BEs in the respective FT, ensuring the completeness of the failure dataset. Our dataset allows us to compare the FT inferred from the dataset with the ground-truth FT, and thereby evaluate the quality of the inferred FT. Experimental setup. Our case studies were executed five times on an E5-2683V4 CPU at 2.10 GHz, with 16 cores supporting 2 threads each on the EEMCS-HPC Cluster of the University of Twente. The evaluation comprises two primary sections: the first, elaborated in Section 4.5.1, concentrates on feature assessment through Principal Component Analysis to discern the most informative features from the CM for inferring FTs. Section 4.5.2 compares the e"cacy of the CM-based metrics with the original FT-MOEA implementation, involving two configurations: FT-MOEA-CM-All includes all 17 features from Table 4.1, and FT-MOEA-CM-Best employs only the top 7 features identified in Section 4.5.1. The evaluation addresses robustness, convergence speed, andscalability. Lastly, we also evaluate FT-MOEA-CM’s features of parallelisation and caching in Section 4.5.3. 4.5 Results 4.5.1 Feature assessment For Step 3 in the FT inference (cf. Figure 4.1), we need to identify the most informative metrics listed in Table 4.1. We conduct this by performing feature assessment by evaluating the importance of di!erent variables in a dataset. Here, the features are the metrics computed from the CM. We use Principal Component Analysis (PCA), a multivariate statistical technique that extracts information in the form of principal components (Abdi and Williams, 2010), which are orthogonal vectors that maximise variance, capturing the most ∗https://gitlab.utwente.nl/fmt/fault-trees/ft-moea †https://github.com/killB0x/ft-moea-cm-server
RkJQdWJsaXNoZXIy MjY0ODMw